[Mesa-dev] [PATCH] anv: Implement VK_ANDROID_native_buffer (v5)

[Chad Versace]

On Mon 25 Sep 2017, Tapani Pälli wrote:
> 
> 
> On 09/22/2017 03:09 AM, Chad Versace wrote:
> > On Thu 21 Sep 2017, Tapani Pälli wrote:
> > > Hi Chad;
> > > 
> > > The build works ok now on Android-IA. There is still something wrong with
> > > 'exec async' though. It behaves differently with small/big apps but
> > > eventually I think it just starts to block .. somewhere. I still need the
> > > big hammer to set device->has_exec_async false to fix that. Please don't
> > > consider that to be a blocker though, we can easily carry such patch in
> > > Android-IA and debug it further.
> > > 
> > > For this patch:
> > > Reviewed-by: Tapani Pälli <tapani.palli at intel.com>
> > 
> > Thanks for testing and review.
> > 
> > How long until you observe the lockup? And does the lockup happen
> > earlier or later for small apps vs big apps? And what apps do you use to
> > reproduce the lock? I'll try to reproduce the lock in my ARC++ setup
> > after returning to the office on Monday (I'm at XDC now).
> > 
> 
> Lockup happens very quickly on any app but some apps manage to do a bit of
> work, here's 3 examples:
> 
> Triangle example from Sascha's demos clears background to blue but then does
> not succeed in rendering triangle.
> 
> RayTracing example in Sascha's demos displays black background for long time
> but then eventually gets VK_ERROR_DEVICE_LOST and crashes as after this it
> still attempts to begin a new render pass.
> 
> VulkanParticleSystem from PowerVR SDK gets VK_ERROR_DEVICE_LOST very early
> with texture uploads and submitting command buffers. Eventually
> ActivityManager kills it because input dispatching timed out.

Tapani, I finally succeeded in building and installing the Sascha demos
onto ARC++ x86_64. I've tested 'triangle', 'gears', 'raytracing', and
'dynamicuniformbuffers', and see no issues. I run each app for about 10
minutes before closing it. Just in case you wanted to test my APKs on
your system, I've uploaded them to
http://kiwitree.net/~chadv/tmp/sascha-demos/.

Maybe the problem is due to triple-vs-double buffering? I believe
SurfaceFlinger in ARC++ uses triple-buffering (that is, the BufferQueue
is 3 deep), but default AOSP uses double-buffering.